Is torch.square(x) faster than torch.pow(x, 2)?

I found several examples of code where torch.pow(x, 2) is used to square the components of a tensor. Would it be more performant to replace such instances with torch.square(x)?

I have not yet come up with a benchmark to test this; I was hoping someone with more knowledge of torch's internals would be able to expand on why one is faster than the other (or not).

1 Answer

In my experiments using timeit for 1M tries to compute a square of a 4096 vector, I found:

1. * fastest with 8.55s
2. square with 12.51s
3. pow with 14.11s